Privacy Preserving Machine Learning Using Homomorphic Encryption

Abstract

Neural network-based machine learning algorithms have shown outstanding results and are currently being widely used in numerous fields. These machine learning algorithms demands considerable computing power for internal calculations and training with big datasets in a reasonable amount of time. In recent years, clouds provide services to facilitate this process, but it introduces new security threats, as the machine learning algorithms mainly rely on the utilization of personal data for training and classification which frequently has privacy implications. To overcome this problem, we propose new approach for operating deep neural networks on encrypted data. Homomorphic encryption is a cryptographic technique, which allows to perform computations on encrypted data, but it also has some limitations associated with it. However, it only supports limited number of addition and multiplication operations in encrypted domain. Existing works only cater simple machine learning algorithms like binary classifiers and simple neural networks in the encrypted domain. Moreover, these simple machine learning algorithms does not provide the required accuracies and also handle a limited number of datasets. To address these issues deeper neural networks are required, which on the other hand increases the computational complexity. In this study, we create novel methods for implementing deep neural networks within the realistic limitations of homomorphic encryption techniques. We mainly concentrate on convolutional neural networks for training and encrypted classification. To begin, we provide techniques for approximating the activation functions typically employed in CNNs (e.g., ReLU and Sigmoid) with low degree polynomials, which is required for efficient homomorphic encryption schemes. The models are then trained using approximation polynomials rather than the original activation functions, and their performance is evaluated. In the end, we apply convolutional neural networks to encrypted data for privacy preserving classification by varying the various Homomorphic encryption scheme’s parameters and evaluate the model performance. The proposed scheme ensures privacy while attaining the maximum accuracy.